Imports

Load Image

Visualize Annotations

Resampling Image

Downsampling Image from (3000, 4000) to (900, 1200)to speed up processing. (Might remove later when the entire pipeline is finished)

Resize Annotations

(Might remove later when entire pipeline is finished.)

Train and Test Split

With Annotations Visualization

Feature Selection

Color, texture, and entropy are chosen as the features for pixel-level classification.

Color

Color is undoubtedly the most revealing feature for trees (Yang et al., 2009). The RGB channels of each aerial image will be converted to CIE L*a*b* color space for a better perceptual uniformity, and to the illumination-invariant color space (Chong et al., 2008) for some robustness against the change of lighting condition. The two color representations will be concatenated to form a 6 dimensional feature vector at each pixel.

CIE L*a*b* Color Space

Illumination Invariant Color Space

Texture

The texture pattern formed by tree leaves often distinguishes trees from similarly colored objects (Yang et al., 2009) such as grass and bushes. The texture feature is generated as a set of filter responses at each pixel by convolving the L channel of each aerial image with a filter-bank. Gaussian derivative filter were empirically chosen to from the filter-bank. Each Gaussian derivative filter is a second derivate of gaussian in the X and Y direction. The filter-bank consists of filters on 3 scales (with σ = 1, √2, 2) and 6 orientations uniformly sampled in [0, π), which generates an 18 dimensional feature vector at each pixel.

Entropy

Entropy measures the uncertainty of a random variable (Yang et al., 2009), which in this case the L channel of aerial images. This helps to differentiate between tree leaves and ground. The entropy of each pixel is computed within 5 x 5, 9 x 9, and 17 x 17 search windows on the L channel of the image. Concatenating the entropy values forms a 3 dimensional feature vector at each pixel.

Visualizing Entropy

Input Features

Finally, by concatenating color, texture and entropy features, a 27 dimensional feature vector is formed at each pixel.

Add Y Values to data set

Training Data Set

Pixel-level Classification

Classification Refinement

Tree Localization

Examine Template Radiuses

Create Templates

Assign P Channel to target image

Template Matching

Apply To Test Image

Color Feature

Texture Feature

Entropy Feature

Concatenate Features

Pixel-level Classification

Refined Classification

Tree Localization

Results

Ground and Result Comparison

True Positives

False Positives

False Negatives

Combined Results

Accuracy

Dump Session

Loading Dumped Session